Compute Library
 20.08
NEFullyConnectedLayer.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2017-2020 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
25 
32 
33 #include <algorithm>
34 #include <cmath>
35 
36 namespace arm_compute
37 {
39 
40 namespace
41 {
42 // Get min, max bound of a quantized assymetric output tensor, with the effect of fused activation
43 std::pair<PixelValue, PixelValue> get_quantized_asymmetric_output_min_max(const QuantizationInfo &q_info, const ActivationLayerInfo &act_info, DataType data_type)
44 {
45  PixelValue type_min{};
46  PixelValue type_max{};
47  std::tie(type_min, type_max) = get_min_max(data_type);
48  const UniformQuantizationInfo q_unif = q_info.uniform();
49 
50  if(act_info.enabled())
51  {
52  switch(act_info.activation())
53  {
55  type_min = PixelValue(q_unif.offset);
56  break;
58  type_min = PixelValue(q_unif.offset);
59  type_max = PixelValue(act_info.a(), data_type, q_info);
60  break;
62  type_min = PixelValue(act_info.b(), data_type, q_info);
63  type_max = PixelValue(act_info.a(), data_type, q_info);
64  break;
65  default:
66  ARM_COMPUTE_ERROR("Activation function not supported.");
67  break;
68  }
69  }
70 
71  return std::make_pair(type_min, type_max);
72 }
73 
74 Status get_gemmlowp_output_stage_info(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, const ActivationLayerInfo &act,
75  GEMMLowpOutputStageInfo &gemmlowp_output_stage_info)
76 {
77  const auto data_type = input->data_type();
78  const QuantizationInfo oq_info = output->quantization_info();
79  const UniformQuantizationInfo iq_unif = input->quantization_info().uniform();
80  const UniformQuantizationInfo wq_unif = weights->quantization_info().uniform();
81  const UniformQuantizationInfo oq_unif = oq_info.uniform();
82 
83  float multiplier = (iq_unif.scale * wq_unif.scale) / oq_unif.scale;
84  int32_t output_multiplier;
85  int32_t output_shift;
86 
87  ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift));
88 
89  PixelValue type_min{};
90  PixelValue type_max{};
91  std::tie(type_min, type_max) = get_quantized_asymmetric_output_min_max(oq_info, act, data_type);
92 
93  gemmlowp_output_stage_info.gemmlowp_multiplier = output_multiplier;
94  gemmlowp_output_stage_info.gemmlowp_shift = output_shift;
95  gemmlowp_output_stage_info.gemmlowp_offset = oq_unif.offset;
96  gemmlowp_output_stage_info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
97  gemmlowp_output_stage_info.gemmlowp_min_bound = type_min.get<int32_t>();
98  gemmlowp_output_stage_info.gemmlowp_max_bound = type_max.get<int32_t>();
99 
100  return Status{};
101 }
102 
103 Status validate_mm(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const ActivationLayerInfo &act)
104 {
105  if(is_data_type_quantized_asymmetric(input->data_type()))
106  {
107  // Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
108  // Extract and negate input and weights offset
109  const QuantizationInfo input_quantization_info(input->quantization_info().uniform().scale, -input->quantization_info().uniform().offset);
110  const QuantizationInfo weights_quantization_info(weights->quantization_info().uniform().scale, -weights->quantization_info().uniform().offset);
111 
112  GEMMLowpOutputStageInfo gemmlowp_output_stage_info;
113  ARM_COMPUTE_RETURN_ON_ERROR(get_gemmlowp_output_stage_info(input, weights, output, act, gemmlowp_output_stage_info));
114 
115  GEMMInfo gemm_info;
116  gemm_info.set_gemmlowp_output_stage(gemmlowp_output_stage_info);
117 
118  // Validate gemmlowp function
119  ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixMultiplyCore::validate(&input->clone()->set_quantization_info(input_quantization_info),
120  &weights->clone()->set_quantization_info(weights_quantization_info),
121  biases,
122  output,
123  gemm_info));
124  }
125  else
126  {
127  ARM_COMPUTE_RETURN_ON_ERROR(NEGEMM::validate(input, weights, biases, output, 1.f, 1.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run */)));
128  }
129 
130  return Status{};
131 }
132 } // namespace
133 
135 {
136  auto k = arm_compute::support::cpp14::make_unique<NETransposeKernel>();
137  k->configure(input, output);
138  _kernel = std::move(k);
139 }
140 
142 {
143  return NETransposeKernel::validate(input, output);
144 }
145 
146 NEFullyConnectedLayer::NEFullyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager, IWeightsManager *weights_manager)
147  : _memory_group(std::move(memory_manager)), _weights_manager(weights_manager), _flatten_kernel(), _convert_weights(), _convert_weights_managed(), _reshape_weights_function(),
148  _reshape_weights_managed_function(), _mm_gemm(nullptr, weights_manager), _mm_gemmlowp(nullptr, weights_manager), _flatten_output(), _converted_weights_output(), _reshape_weights_output(),
149  _original_weights(nullptr), _are_weights_converted(true), _are_weights_reshaped(false), _is_fc_after_conv(false), _is_quantized_asymmetric(false), _is_prepared(false)
150 {
151 }
152 
153 void NEFullyConnectedLayer::configure_mm(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const ActivationLayerInfo &act)
154 {
155  if(_is_quantized_asymmetric)
156  {
157  // Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
158  // Extract and negate input and weights offset
159  const QuantizationInfo input_quantization_info = input->info()->quantization_info();
161 
162  input->info()->set_quantization_info(QuantizationInfo(input_quantization_info.uniform().scale, -input_quantization_info.uniform().offset));
164 
165  // Configure gemmlowp function and output stage for asymmetric quantized types
166  GEMMLowpOutputStageInfo gemmlowp_output_stage_info;
167  const Status status = get_gemmlowp_output_stage_info(input->info(), weights->info(), output->info(), act, gemmlowp_output_stage_info);
169 
170  GEMMInfo gemm_info;
171  gemm_info.set_gemmlowp_output_stage(gemmlowp_output_stage_info);
172  gemm_info.set_activation_info(act);
173  _mm_gemmlowp.configure(input, weights, biases, output, gemm_info);
174 
175  // Revert back QuantizatioInfo as input and weights could be used in other fully connected layers
176  input->info()->set_quantization_info(input_quantization_info);
178  }
179  else
180  {
181  // Configure matrix multiply kernel
182  GEMMInfo gemm_info(false, false, true /* Reshape weights only for the first run */);
183  gemm_info.set_activation_info(act);
184  _mm_gemm.configure(input, weights, biases, output, 1.f, 1.0f, gemm_info);
185  }
186 }
187 
188 void NEFullyConnectedLayer::configure_conv_fc(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const ActivationLayerInfo &act)
189 {
190  ARM_COMPUTE_ERROR_ON((weights->info()->dimension(1) != (input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2))));
191 
192  // If the fully connected layer is called after a convolution layer, the input tensor must be linearized
193 
194  // Initialize output tensor for flatten
195  TensorShape shape_flatten = compute_flatten_shape(input->info());
196  _flatten_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_flatten));
197 
198  // Configure flatten kernel
199  _memory_group.manage(&_flatten_output);
200  _flatten_kernel.configure(input, &_flatten_output);
201 
202  // Configure matrix multiply kernel
203  configure_mm(&_flatten_output, weights, biases, output, act);
204 
205  // Allocate the output tensor for flatten once all the configure methods have been called
206  _flatten_output.allocator()->allocate();
207 }
208 
209 void NEFullyConnectedLayer::configure_fc_fc(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const ActivationLayerInfo &act)
210 {
211  ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != weights->info()->dimension(1));
212 
213  // Configure matrix multiply kernel
214  configure_mm(input, weights, biases, output, act);
215 }
216 
217 void NEFullyConnectedLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output,
218  FullyConnectedLayerInfo fc_info)
219 {
220  // Perform validate step
223  weights->info(),
224  biases != nullptr ? biases->info() : nullptr,
225  output->info(),
226  fc_info));
227 
228  _are_weights_converted = true;
229  _are_weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
230  _is_fc_after_conv = true;
231  _is_quantized_asymmetric = is_data_type_quantized_asymmetric(input->info()->data_type());
232  _original_weights = weights;
233 
234  if(_weights_manager)
235  {
236  _weights_manager->manage(weights);
237  }
238 
239  // With the Fully Connected layer we can have 4 different cases:
240  // 1) Convolution layer -> Fully Connected layer without batches
241  // 2) Fully Connected layer -> Fully Connected layer without batches
242  // 3) Convolution layer -> Fully Connected layer with batches
243  // 4) Fully Connected layer -> Fully Connected layer with batches
244 
245  const ITensor *weights_to_use = weights;
246 
247  // Check if we have a fully connected layer with batches
248  const bool is_batched_fc_layer = output->info()->dimension(1) > 1;
249  if(is_batched_fc_layer)
250  {
251  _is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(input->info()->tensor_shape().cbegin() + 3,
252  input->info()->tensor_shape().cend(),
253  output->info()->tensor_shape().cbegin() + 1));
254  }
255  else
256  {
257  _is_fc_after_conv = input->info()->num_dimensions() > 1;
258  }
259 
260  // Reshape weights if needed
261  if(!_are_weights_reshaped)
262  {
263  if(_weights_manager && _weights_manager->are_weights_managed(weights))
264  {
265  _reshape_weights_managed_function.configure(weights);
266  weights_to_use = _weights_manager->acquire(weights, &_reshape_weights_managed_function);
267  }
268  else
269  {
270  // Reshape the weights
271  _reshape_weights_function.configure(weights, &_reshape_weights_output);
272  weights_to_use = &_reshape_weights_output;
273  }
274  }
275 
276  // Convert weights if needed
277  if(_is_fc_after_conv && (input->info()->data_layout() != fc_info.weights_trained_layout))
278  {
279  if(_weights_manager && _weights_manager->are_weights_managed(weights_to_use))
280  {
281  _convert_weights_managed.configure(weights_to_use,
282  input->info()->tensor_shape(),
283  fc_info.weights_trained_layout);
284  weights_to_use = _weights_manager->acquire(weights, &_convert_weights_managed);
285  }
286  else
287  {
288  // Convert weights
289  _convert_weights.configure(weights_to_use,
290  &_converted_weights_output,
291  input->info()->tensor_shape(),
292  fc_info.weights_trained_layout);
293 
294  weights_to_use = &_converted_weights_output;
295  }
296  _are_weights_converted = false;
297  }
298 
299  if(_is_fc_after_conv)
300  {
301  // Fully Connected layer after a Convolution Layer without batches
302  configure_conv_fc(input, weights_to_use, biases, output, fc_info.activation_info);
303  }
304  else
305  {
306  // Fully Connected layer after a Fully Connected Layer without batches
307  configure_fc_fc(input, weights_to_use, biases, output, fc_info.activation_info);
308  }
309 
310  _are_weights_reshaped = _are_weights_reshaped || fc_info.retain_internal_weights;
311 }
312 
314  FullyConnectedLayerInfo fc_info)
315 {
320  ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 2);
321  ARM_COMPUTE_RETURN_ERROR_ON(biases != nullptr && biases->num_dimensions() > 1);
322 
323  bool weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
324  bool is_fc_after_conv = true;
325 
326  const ITensorInfo &flatten_input = TensorInfo(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_flatten_shape(input)));
327  const ITensorInfo &reshaped_weights = TensorInfo(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_transposed_shape(*weights)));
328  const ITensorInfo &converted_weights = weights_reshaped ? TensorInfo(weights->clone()->set_is_resizable(true).reset_padding()) : TensorInfo(*reshaped_weights.clone());
329 
330  // With the Fully Connected layer we can have 4 different cases:
331  // 1) Convolution layer -> Fully Connected layer without batches
332  // 2) Fully Connected layer -> Fully Connected layer without batches
333  // 3) Convolution layer -> Fully Connected layer with batches
334  // 4) Fully Connected layer -> Fully Connected layer with batches
335 
336  const ITensorInfo *input_to_use = input;
337  const ITensorInfo *weights_to_use = weights;
338 
339  // Check if we have a fully connected layer with batches
340  const bool is_batched_fc_layer = output->dimension(1) > 1;
341 
342  if(is_batched_fc_layer)
343  {
344  is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(input->tensor_shape().cbegin() + 3,
345  input->tensor_shape().cend(),
346  output->tensor_shape().cbegin() + 1));
347  }
348  else
349  {
350  is_fc_after_conv = input->num_dimensions() > 1;
351  }
352 
353  if(!weights_reshaped)
354  {
355  // Validate reshape weights kernel
357  weights_to_use = &reshaped_weights;
358  }
359 
360  if(is_fc_after_conv && (input->data_layout() != fc_info.weights_trained_layout))
361  {
362  // Validate convert weights kernel
364  &converted_weights,
365  input->tensor_shape(),
366  fc_info.weights_trained_layout));
367  weights_to_use = &converted_weights;
368  }
369 
370  if(is_fc_after_conv)
371  {
372  // Fully Connected layer after a Convolution Layer without batches
373  ARM_COMPUTE_RETURN_ERROR_ON((weights_to_use->dimension(1) != (input->dimension(0) * input->dimension(1) * input->dimension(2))));
374 
375  // Validate flatten kernel
377  input_to_use = &flatten_input;
378  }
379  else
380  {
381  // Fully Connected layer after a Fully Connected Layer without batches
382  ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != weights_to_use->dimension(1));
383  }
384  // Validate matrix multiply kernel
385  ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(input_to_use, weights_to_use, biases, output, fc_info.activation_info));
386 
387  return Status{};
388 }
389 
391 {
392  prepare();
393 
394  MemoryGroupResourceScope scope_mg(_memory_group);
395 
396  // Linearize input if it comes from a convolutional layer
397  if(_is_fc_after_conv)
398  {
399  NEScheduler::get().schedule(&_flatten_kernel, Window::DimY);
400  }
401 
402  // Run matrix multiply
403  if(_is_quantized_asymmetric)
404  {
405  _mm_gemmlowp.run();
406  }
407  else
408  {
409  _mm_gemm.run();
410  }
411 }
412 
414 {
415  if(!_is_prepared)
416  {
417  if(!_weights_manager)
418  {
419  ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
420  }
421 
422  auto release_unused = [](Tensor * w)
423  {
424  if(!w->is_used())
425  {
426  w->allocator()->free();
427  }
428  };
429 
430  // Pointer to current weights
431  const ITensor *cur_weights = _original_weights;
432 
433  // Reshape of the weights (happens only once)
434  if(!_are_weights_reshaped)
435  {
436  if(_weights_manager && _weights_manager->are_weights_managed(_original_weights))
437  {
438  cur_weights = _weights_manager->run(cur_weights, &_reshape_weights_managed_function);
439  }
440  else
441  {
442  // Reshape of the weights (happens only once)
443  if(!_are_weights_reshaped)
444  {
445  // Run reshape weights kernel and mark weights as unused
446  _reshape_weights_output.allocator()->allocate();
447  _reshape_weights_function.run();
448  }
449  cur_weights->mark_as_unused();
450  cur_weights = &_reshape_weights_output;
451  }
452  _are_weights_reshaped = true;
453  }
454 
455  // Convert weights if needed (happens only once)
456  if(!_are_weights_converted)
457  {
458  if(_weights_manager && _weights_manager->are_weights_managed(cur_weights))
459  {
460  _weights_manager->run(cur_weights, &_convert_weights_managed);
461  }
462  else
463  {
464  _converted_weights_output.allocator()->allocate();
465  _convert_weights.run();
466  cur_weights->mark_as_unused();
467  }
468 
469  _are_weights_converted = true;
470  }
471 
472  // Release reshaped weights if unused
473  release_unused(&_reshape_weights_output);
474 
475  // Prepare GEMM prepare and release unused weights
476  if(!_is_quantized_asymmetric)
477  {
478  _mm_gemm.prepare();
479  }
480 
481  // Release converted weights if unused
482  release_unused(&_reshape_weights_output);
483  release_unused(&_converted_weights_output);
484 
485  _is_prepared = true;
486  }
487 }
488 } // namespace arm_compute
virtual size_t num_dimensions() const =0
The number of dimensions of the tensor (rank)
SimpleTensor< float > w
Definition: DFT.cpp:156
Quantize using a fixed point multiplication.
void run() override final
Run the kernels contained in the function.
void init(const TensorAllocator &allocator, const Coordinates &coords, TensorInfo &sub_info)
Shares the same backing memory with another tensor allocator, while the tensor info might be differen...
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
bool retain_internal_weights
Retain internal reshaped weights.
Definition: Types.h:1585
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
Definition: Validate.h:545
#define ARM_COMPUTE_ERROR(msg)
Print the given message then throw an std::runtime_error.
Definition: Error.h:352
#define ARM_COMPUTE_RETURN_ON_ERROR(status)
Checks if a status contains an error and returns it.
Definition: Error.h:204
bool is_used() const
Flags if the tensor is used or not.
Definition: ITensor.cpp:163
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:792
1 channel, 1 F32 per channel
#define ARM_COMPUTE_ERROR_ON(cond)
If the condition is true then an error message is printed and an exception thrown.
Definition: Error.h:466
Fully connected layer info.
Definition: Types.h:1580
Store the tensor's metadata.
Definition: ITensorInfo.h:40
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
Status calculate_quantized_multiplier(float multiplier, int32_t *quant_multiplier, int32_t *shift, bool ignore_epsilon=false)
Calculate quantized representation of multiplier.
void manage(const ITensor *weights, ITransformWeights *parent=nullptr)
Start managing a weights tensor.
Status class.
Definition: Error.h:52
#define ARM_COMPUTE_RETURN_ERROR_ON(cond)
If the condition is true, an error is returned.
Definition: Error.h:296
Activation Layer Information class.
Definition: Types.h:1517
Interface for NEON tensor.
Definition: ITensor.h:36
Copyright (c) 2017-2020 Arm Limited.
1 channel, 1 F16 per channel
TensorAllocator * allocator()
Return a pointer to the tensor's allocator.
Definition: Tensor.cpp:48
ITensorInfo * info() const override
Interface to be implemented by the child class to return the tensor's metadata.
Definition: Tensor.cpp:33
TensorShape compute_transposed_shape(const ITensorInfo &input)
Calculate the transposed shape of a tensor.
void mark_as_unused() const
Marks a tensor as unused.
Definition: ITensor.cpp:168
static Status validate(const ITensorInfo *input, const ITensorInfo *output)
Static function to check if given info will lead to a valid configuration of NEFullyConnectedLayerRes...
void manage(IMemoryManageable *obj) override
Sets a object to be managed by the given memory group.
Definition: MemoryGroup.h:79
bool are_weights_managed(const ITensor *weights)
Check if the weights are managed.
TensorShape compute_flatten_shape(const ITensorInfo *input)
Calculate the flattened output shape of a tensor.
static Status validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info=GEMMInfo())
Static function to check if given info will lead to a valid configuration of NEGEMM.
Definition: NEGEMM.cpp:163
Quantization information.
void run() override
Run the kernels contained in the function.
Definition: NEGEMM.cpp:281
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
Definition: Error.h:152
void configure(const ITensor *input, ITensor *output)
Set the input and output of the kernel.
void configure(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *output, const GEMMInfo &gemm_info=GEMMInfo())
Initialise the kernel's inputs, output.
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
void run() override
Run the kernels contained in the function.
quantized, asymmetric fixed-point 8-bit number unsigned
bool are_weights_reshaped
Reshape the weights tensor if false.
Definition: Types.h:1584
void configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, FullyConnectedLayerInfo fc_info=FullyConnectedLayerInfo())
Set the input and output tensors.
void allocate() override
Allocate size specified by TensorInfo of CPU memory.
NEFullyConnectedLayer(std::shared_ptr< IMemoryManager > memory_manager=nullptr, IWeightsManager *weights_manager=nullptr)
Constructor.
UniformQuantizationInfo uniform() const
Return per layer quantization info.
GEMMLowp output stage info.
Definition: Types.h:1881
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor's metadata.
void configure(const ITensor *input, const TensorShape &original_input_shape, DataLayout data_layout)
Basic implementation of the tensor interface.
Definition: Tensor.h:37
virtual ITensorInfo & set_quantization_info(const QuantizationInfo &quantization_info)=0
Set the quantization settings (scale and offset) of the tensor.
static Status validate(const ITensorInfo *input, const ITensorInfo *output)
Static function to check if given info will lead to a valid configuration of NETransposeKernel.
ActivationLayerInfo activation_info
Fused activation to apply after the matrix multiplication.
Definition: Types.h:1587
Weights manager interface to handle weights transformations.
virtual QuantizationInfo quantization_info() const =0
Get the quantization settings (scale and offset) of the tensor.
void configure(const ITensor *input, ITensor *output, const TensorShape &original_input_shape, DataLayout data_layout)
Initialize the function.
bool is_data_type_quantized_asymmetric(DataType dt)
Check if a given data type is of asymmetric quantized type.
Definition: Utils.h:1143
__constant DATA_TYPE16 type_min
Definition: minmaxloc.cl:46
#define ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(...)
Definition: Validate.h:163
static constexpr size_t DimY
Alias for dimension 1 also known as Y dimension.
Definition: Window.h:45
std::array< T, num_max_dimensions >::const_iterator cbegin() const
Returns a read-only (constant) iterator that points to the first element in the dimension array.
Definition: Dimensions.h:210
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:161
Memory group resources scope handling class.
Definition: IMemoryGroup.h:82
void set_gemmlowp_output_stage(GEMMLowpOutputStageInfo &output_stage)
Sets GEMMLowp output stage.
Definition: Types.h:2043
virtual void schedule(ICPPKernel *kernel, const Hints &hints)=0
Runs the kernel in the same thread as the caller synchronously.
void run() override
Run the kernels contained in the function.
static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, FullyConnectedLayerInfo fc_info=FullyConnectedLayerInfo())
Static function to check if given info will lead to a valid configuration of NEFullyConnectedLayer.
DataLayout weights_trained_layout
Layout that the weights have been trained with.
Definition: Types.h:1582
static Status validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape, DataLayout data_layout)
Static function to check if given info will lead to a valid configuration of NEConvertFullyConnectedW...
void configure(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *d, float alpha, float beta, const GEMMInfo &gemm_info=GEMMInfo())
Initialise the kernel's inputs, output.
Definition: NEGEMM.cpp:51
void prepare() override
Prepare the function for executing.
Definition: NEGEMM.cpp:331
const QuantizationInfo weights_quantization_info
__constant DATA_TYPE16 type_max
Definition: minmaxloc.cl:47
bool transpose_weights
Transpose weights if true.
Definition: Types.h:1583
void configure(const ITensor *input, ITensor *output)
Set the input and output tensors.
Store the tensor's metadata.
Definition: TensorInfo.h:45
static Status validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, const GEMMInfo &gemm_info=GEMMInfo())
Static function to check if given info will lead to a valid configuration of NEGEMMLowpMatrixMultiply...
GEMM information class.
Definition: Types.h:1932
ITensor * run(const ITensor *weights, ITransformWeights *weights_transform)
Run the reshape function.
quantized, asymmetric fixed-point 8-bit number signed
void prepare() override
Prepare the function for executing.
static Status validate(const ITensorInfo *input, const ITensorInfo *output)
Static function to check if given info will lead to a valid configuration of NEFlattenLayerKernel.
static constexpr size_t num_max_dimensions
Number of dimensions the tensor has.
Definition: Dimensions.h:45
DataType
Available data types.
Definition: Types.h:77
std::tuple< PixelValue, PixelValue > get_min_max(DataType dt)
Compute the mininum and maximum values a data type can take.
Definition: Utils.h:560
ErrorCode error_code() const
Gets error code.
Definition: Error.h:89
ITensor * acquire(const ITensor *weights, ITransformWeights *weights_transform)
Acquire the requested reshape tensor of the selected weights.
void run() override
Run the kernels contained in the function.
static IScheduler & get()
Access the scheduler singleton.
Definition: Scheduler.cpp:95